CN103201707A - System and method for inputting text into electronic devices - Google Patents

Abstract

A text prediction engine, a system comprising a text prediction engine, and a method for generating sequence predictions. The text prediction engine, system and method generate a final set of sequence predictions, each with an associated probability value.

Description

Be used for text prediction engine, system and method to the electronic equipment input text

Technical field

The present invention relates generally to a kind of for text prediction engine, system and method to the electronic equipment input text.

Background technology

Some existing inventions utilize multiple different technologies that the improvement method of electronic equipment user version input is provided.Yet well-known, disclosed related system at first is faced with the problem of using stable and fully integrated probability model predictive user expection to write text.

Summary of the invention

In a first aspect of the present invention, a kind of text prediction engine is provided, comprising: at least one model, it is used for having first group of sequence that dependent probability is estimated from the generation of evidence source; The probability maker, it is used for receiving the described first group of sequence that has the dependent probability appraisal and generates one group of sequence prediction that has the dependent probability value; Wherein, received under the situation of all possible sequence the described probable value of normalization on all possible sequence prediction that is generated by described probability maker by described probability maker given.

Preferably, described text prediction engine comprises priority model, and it is used for generating the second group of sequence that has the dependent probability appraisal.

Preferably, described model generates first group of sequence according to the uncertainty in described evidence source and the described evidence source.Preferably, described probability maker is used for receiving described first, second group sequence that has the dependent probability appraisal.

Described probability maker preferably by n most probable sequence prediction represented the constant addition with the probable value that remains possible sequence prediction, is estimated the normalization factor of described probable value.Described constant is represented the probable value by the remaining possibility sequence prediction of described model and the generation of described priority model.

Described model comprises a plurality of models that have a plurality of first group of sequence of dependent probability appraisal for generation.In one embodiment, described a plurality of model generates a plurality of first group of sequence according to a plurality of evidences source.

Preferably, described text prediction engine is the part of a certain system, and described user input text by one or more users select, character input or speech recognition be input in this system.

Described text prediction engine comprises given linguistic context sequence according to corresponding model probability is weighted the probable value of described sequence prediction.In one embodiment, described a plurality of model comprises a plurality of language models corresponding with multiple different language; And the probable value of the described text prediction engine pair sequence prediction corresponding with the language model of the most probable language that relates to user input text is carried out five-star weighting.

Each evidence source is moulded by the corresponding model that is used for generating the sequence that has the dependent probability appraisal.Under the situation of given described sequence prediction, described probability maker preferably with each evidence source as other on evidence the separate component of having ready conditions in source handle.

In a preferred embodiment of described text prediction engine, described model comprises context model and input model, and described context model is used for receiving the text of user's input and generates one group of sequence with described input model to be estimated with relevant probability; And described priority model comprises for generating the target priority model that one group of sequence and dependent probability are estimated.Described input model preferably includes candidate's model and language model.Described context model preferably includes candidate's model and prefix matching model.Described target priority model preferably includes character model and a meta-model.

In a second aspect of the present invention, a kind of system is provided, comprising: user interface, it is used for receiving the text by user's input; Text prediction engine, it be used for to receive from the described text of described user interface input and generates one group of sequence prediction that has the dependent probability value, wherein, all possible sequence prediction on the described probable value of normalization; Wherein, described text prediction engine also is used for providing described sequence prediction to described user interface.

Preferably, described input model comprises candidate's model and language model.Preferably, described context model comprises candidate's model and prefix matching model.Preferably, described target priority model comprises character model and a meta-model.

In a third aspect of the present invention, a kind of method of process user text input is provided, comprising: receive the text that inputs to user interface; Use text prediction engine to generate one group of sequence prediction and relevant probable value, wherein, the described probable value of normalization on all possible sequence prediction; Described sequence prediction is offered described user interface.

The step that generates the normalization probable value preferably includes: by the probable value of n most probable sequence prediction is represented the constant addition with remaining probable value that may sequence prediction, estimate the normalization factor of described probable value.

This method also comprises: described sequence prediction is presented on the described user interface selects for the user.Preferably, by described text prediction engine described sequence prediction is sorted, show in order for described user interface.Only when the corresponding probable value of described sequence prediction during more than or equal to first threshold values, described sequence prediction is offered described user interface.Similarly, only when the corresponding probable value of described sequence prediction during more than or equal to first threshold values, said system offers described user interface with described sequence prediction.

Preferably, at least one in the described sequence prediction is equivalent to be inputed to by the user adjustment or the invulnerable release of the text of described user interface.

Described method also comprises: input has the sequence prediction greater than second threshold values or the probable value on second threshold values automatically.Similarly, in one embodiment, said system is imported the sequence prediction that has greater than second threshold values or the probable value on second threshold values automatically.

The probability maker that uses in this method preferably includes for generating a plurality of models that the corresponding model of one group of sequence prediction and dependent probability value and basis comprises the described probable value of probability weight of given linguistic context sequence.

The present invention also provides a kind of computer program, comprising: computer-readable medium wherein stores be used to making processor carry out the computer program of said method.

The invention still further relates to a kind of text prediction engine for the formation sequence prediction and be used for the formation sequence prediction for demonstration and user selected system and method.In one embodiment, the present invention relates to correct mistakes automatically list entries system and realize the method for this correction.In a preferred embodiment, the invention provides a text prediction engine and generate the one group of ultimate sequence prediction that has the dependent probability value by estimating in conjunction with the different probability of an arbitrarily sequence expection.Text prediction engine of the present invention, system and method can provide the prediction based on any corroboration source thus.Can be by distributing correct probability to each expected sequence but not be the sequence rank, realization this purpose.By distributing correct probable value, but analysis of allocated is given the evolution of the probability of different entries, and can compare given entry on two different time points or the probability of one group of entry.This means in particular prediction under given pre-set threshold value " trust " situation (confidence), can use pre-set threshold value to come the behavior of regulating system.For instance, only show the sequence predict, if perhaps the accuracy of system estimation probability surpasses 0.75, or the sequence that in other words dopes have at least 75% may be accurately, revise this moment automatically.If use certain special evaluation to come to be the element rank, such as those values that can't reliably compare between the sequence on the time point, then this reasoning is irrealizable.

In order to generate correct probable value, the present invention preferably provide the normalization of a kind of effective estimation all sequences to add and method.

Below with reference to following accompanying drawing, introduce the present invention in detail.

Description of drawings

Fig. 1 is the synoptic diagram of the senior predict of the present invention;

Fig. 2 is the synoptic diagram of the preferred predict example of the present invention.

Embodiment

Definition:

● the specification unit symbol that character-expression is basic;

● the finite aggregate of character set-character;

● sequence-sorted finite length character string;

● if the initial character of prefix-originate in each sequence is identical, and has continuous mapping one by one and length (s)≤length (s'), and then sequence s is the prefix of another sequence s ';

● proper prefix-as infructescence s is the prefix of sequence s ', and length (s)＜length (s'), and then sequence s is the proper prefix of another sequence s ';

● language-a group (being generally limited) is specifically write or the order character of oral statement;

● text-from one or more language, extract write data;

● system-theme of the present invention;

● the user-with the mutual personnel of said system.

Generally but not unique, can implement system of the present invention referring to Fig. 1.Fig. 1 is the calcspar of advanced text predict of the present invention.Native system comprises one group of most probable sequence prediction S that expects input by the user of generation ^FText prediction engine.Each sequence prediction has a relative probable value.

As shown in Figure 1, text prediction engine preferably includes being used for from a plurality of evidences source e of a plurality of training ₁, e ₂, e ₃, e ₄Deng in make probability inference model M ₁, M ₂, M ₃And probability maker (PG, probability generator).Yet, in other embodiments, a model of training and an evidence source can be arranged.

The potential evidence source e of more existing any types ₁, e ₂Deng.The example in these evidence sources comprises:

● the sequence that the user has imported;

● the current entry/phrase imported of user;

● the historical series by user's input of storage;

● user's mother tongue;

● the specific type language of input;

● positive input has the application of current sequence;

● the target message recipient in message transmission environment;

● time/date;

● as the position of the equipment of native system main frame.

Universal model

The target of native system is, expects that according to the user possibility of a certain sequence of input is the sequence rank in the given language subset.In Probability, this is equivalent to grade for the sequence in the S set, and this set is arranged by following expression formula:

P(s∈S|e,M) （1）

Wherein, e is the evidence that observes, and the trained model set of M for being used for making probability inference.In other words, native system will be in the set of all sequences that can extract prediction the evaluation condition probability.Target sequence is expressed as s.

In order to simplify the anabolic process of the prediction that comes from the different pieces of information source, in a preferred embodiment, target sequence s is defined as the prediction that comes from particular source.

Each model among the M is trained on particular source.Therefore, can be by the model representation particular source among the M, and the S set in the expression formula (1) relates to all different entries (or sequence) that generated by the model among the M.Provide the prediction entry by interrogation model.This entry is relevant with its source model.Because this entry is relevant with its source model, so it is different from and comes from other models but the entry of morphology unanimity.This relevant can being implied in the above-mentioned data.Yet this entry can be marked with the identifier relevant with this entry source model.

Predict preferably that at this two other the identical prediction that will come from the different pieces of information source is considered as difference in conjunction with in handling.Come from the sequence of different models for combination to obtain predicting list, by removing the prediction of repetition, this sequence is simply sorted.In this preferred operations, given morphology entry/sequence is carried out the most probable assessment, and abandon (least possibility) entry that any morphology repeats.

Be example with a limiting examples, if M comprises two linguistic context language models, French (LM _French) and English (LM _English), entry " pain " may appear in these two models, and occurs twice in S, once be related with the French model, and another time is related with the English model.Like this under the situation of given one group of special evidence (wherein, in this case, this evidence is prediction entry " pain " linguistic context before), entry " pain " carried out the assessment of opening in twice minute.

These assessments relate to two different sequences (is come from the French model, and is come from the English model), but because its morphology is identical, need not all to present to the user.Therefore, according to this preferred embodiment, given morphology sequence is carried out the most probable assessment, and discard any morphology dittograph bar.

In order to expect that according to the user possibility of a certain sequence of input is the sequence rank in the given language subset, need in the calculation expression (1) conditional probability P (s ∈ S|e, M).In order to determine this probability, it is as follows to use Bayesian formula (Bayes ' rule) to rearrange this expression formula:

$\frac{P\left(e\right|s,M)P\left(s\right|M)}{P\left(e\right|M)}---\left(2\right)$

And the target sequence in this expression formula denominator of marginalisation, obtain thus:

$\frac{P\left(e\right|s,M)P\left(s\right|M)}{{\mathrm{\Σ}}_{j=1}^{\left|S\right|}P\left(e\right|s_j,M)P\left(s_j\right|M)}---\left(3\right)$

In a preferred embodiment, (e|s M), supposes: under the situation of given above-mentioned target sequence, evidence can be split into according to correlation model [M in order to calculate P ₁M _N] under the non-overlapped set [e that obtains respectively of a certain distribution ₁E _N].This independence assumption can be write as:

$P\left(e\right|s,M)={\mathrm{\Π}}_{i=1}^N\left[P\left(e_i\right|s,M_i)\right]---\left(4\right)$

And be expressed as follows:

Suppose 1: under the situation of given target sequence, evidence is split as different set, thereby makes the evidence in each set separate conditionally to each other.

Wherein, each e _iHas relative model M _iLike this, can make up a framework, in this framework, can any evidence source be combined in the high mode of counting yield.In a preferred embodiment, model R ∈ M is preferentially relevant with target sequence.Consider this hypothesis, it is as follows that we can explain expression formula (3) again:

$\frac{P\left(s\right|R){\mathrm{\Π}}_{i=1}^{N}P\left(e_i\right|s,M_i)}{{\mathrm{\Σ}}_{j=1}^{\left|S\right|}P\left(s_{j}R\right){\mathrm{\Π}}_{i=1}^{N}P\left(e_i\right|s_j,M_i)}---\left(5\right)$

Therefore, in a preferred embodiment, can be by calculating the preferential P of target sequence (s|R) and each evidential probability P (e _i| s, M _i), calculate the conditional probability of expression formula (1).

Therefore denominator in the expression formula (5) is steady state value with respect to s, can not have influence on rank, and it is the normalization factor of the probable value that calculates or rather.In a preferred embodiment, this steady state value is estimated as subclass and a constant sum of most probable sequence, has to calculate the expression formula 13-15 of S(in vide infra to overcome) in the problem of conditional probability of all sequences.Because the Zipfian(Qi Pufu of some elephants of speaking naturally distributes) characteristic, this method is rationalized, the minority probability event has most probability mass in this characteristic.It is the example that power law distributes that Qi Pufu distributes, and wherein, the frequency of given event and its rank roughly are inversely proportional to.

Expression formula (5) provides and will import the principle method that the different evidences source of purpose combines about text, and at the given e of source on evidence ₁, e ₂... situation under by one group of model R that trained, M ₁, M ₂... generate one group of sequence S _R, S ₁, S ₂... the conditional probability value P relevant with one group _R, P ₁, P ₂... implement optimum decision system of the present invention.Model R is used for calculating priority target sequence probability P (s|R), each model M simultaneously ₁, M ₂... calculate each evidential probability P (e _i| s, M _i).Each model is exported one group of sequence S _iWith one group of relevant conditional probability P _iEach model M ₁, M ₂... comprise one or more submodels.Probability maker PG with sequence and relevant conditional probability as input, and export one group with probable value P _FRelevant ultimate sequence S _FProbability maker PG can combine prediction as above-mentioned preferred process, that is to say, will predict according to probability order rank, and only delete and repeat prediction arbitrarily.With final probable value P _FRelevant sequence S _FCan be presented on the user interface of native system with the form such as tabulation, browse and select for the user.The user can select by making prediction or operation be equipped with the device of native system by other means, carry out alternately with native system, so fresh evidence more.When text inputs to native system, renewable each model R, M ₁M _N

What the invention provides passes through two kinds of method for optimizing that marginalisation is presented on the evidential probability in the evidence candidate lexical or textual analysis calculating probability framework in the graphics frame, although can also use additive method.Below, will introduce this two kinds of method for optimizing.

Candidate's model 1

The probability that comes from the evidence in a certain evidence source in formation is estimated P (e _i| s, M _i) time, help usually to come representation model according to the candidate as the intergrade between ' user's expection ' sequence and the evidence that observes.If come representation model according to the candidate, probability P (e then _i| s, M _i) can be expressed as again:

$P\left(e\right|s,M)={\mathrm{\Σ}}_{j=1}^{K}P\left(e\right|c_j,s,M_{\mathrm{candidate}})P\left(c_j\right|s,M_{\mathrm{sequence}})---\left(6\right)$

Wherein, c _jBe a certain candidate, and have two submodels of the M in given evidence source: candidate's model M _CandidateWith series model M _SequenceCritical assumptions herein are as follows:

Suppose 2: under given candidate's situation, the probability under the above-mentioned model can be expressed as candidate's marginalisation, and wherein, evidence is independent of target sequence conditionally.

Use this hypothesis, can from the evidence entry, delete the dependence for s:

$P\left(e\right|s,M)={\mathrm{\Σ}}_{j=1}^{K}P\left(e\right|c_j,s,M_{\mathrm{candidate}})P\left(c_j\right|s,M_{\mathrm{sequence}})---\left(7\right)$

The attribute of candidate's model can also be encoded into the form of graphical model.This graphical model has been described the relation as follows between variable and the model:

Candidate's model 2

Another variant of candidate's model at first uses Bayesian formula conversion evidential probability:

$P\left(e\right|s,M)=\frac{P\left(s\right|e,M)P\left(e\right|M)}{P\left(s\right|M)}---\left(8\right)$

In one embodiment, evidence condition sequence probability can be expressed as again:

$P\left(s\right|e,M)={\mathrm{\Σ}}_{j=1}^{K}P\left(s\right|c_j,e,M_{\mathrm{sequence}})P\left(c_j\right|e,M_{\mathrm{candidate}})---\left(9\right)$

Wherein, c _jBe a certain candidate, and still have two submodels of the M in given evidence source: candidate's model M _CandidateWith series model M _SequenceLike this, critical assumptions are as follows:

Suppose 3: under given candidate's situation, the probability under the above-mentioned model can be expressed as candidate's marginalisation, and wherein, target sequence is independent of evidence conditionally.

Use this hypothesis, can from the evidence entry, delete the dependence for s:

$P\left(s\right|e,M)={\mathrm{\Σ}}_{j=1}^{K}P\left(s\right|c_j,M_{\mathrm{candidate}})P\left(c_j\right|e,M_{\mathrm{sequence}})---\left(10\right)$

The graphical model of candidate's model of this version is as follows:

And complete evidential probability is:

$P\left(e\right|s,M)=\frac{{\mathrm{\Σ}}_{j=1}^{K}P\left(s\right|c_j,M_{\mathrm{sequence}})P\left(c_j\right|e,M_{\mathrm{candidate}})P\left(e\right|M)}{P\left(s\right|M)}---\left(11\right)$

Particular module

With reference to showing prediction engine from two homologies not: Fig. 2 of the native system preferred embodiment of extraction evidence linguistic context (context) and the input (input), we utilize general candidate's model, have proposed a particular example of native system.Yet as mentioned above, native system is not limited to linguistic context and imports as evidence.If used other or extra evidence source, native system will be correspondingly according to this type of evidence source generation forecast.

Generally, linguistic context is represented the evidence that observes about user's input text, and the clear proof certificate that observes of input text in the ban about the user is then represented in input.For instance, if the user has imported English sequence " My name is B ", then we can think that the linguistic context evidence is sequence " My name is ", and the input evidence is sequence " B ".Yet, only just for example, it should be noted that in prevailing form this model can not made any concrete expression to the special shape of observed evidence.For example, in fact the input evidence may be the touch coordinate that some row come from dummy keyboard.

As shown in Figure 2, evidence (input and linguistic context) is used as the input of prediction engine, in this prediction engine, preferably exist three model R, M _Context, M _Input, each model preferably includes at least two submodels (character model, a meta-model; Candidate's model, language model; Candidate's model, prefix matching model).As shown in Figure 2, this prediction engine preferably includes target sequence priority model R.Although this feature is preferred, native system is not limited to comprise the embodiment of target sequence priority model R.

Target sequence priority model R comprises:

● character model-not fixedly the sequence in the language of vocabulary concept implement to distribute.General using Markov model (Markov model) is implemented this distribution in character string.

Character model is a kind of series model of being set up by the alternative entry of character.For example, if the set in the training is " explaining ", a meta-model may be following appearance:

P(e)=0.1

P(x)=0.1

P(p)=0.1

P(l)=0.1

P(a)=0.1

P(i)=0.2

P(n)=0.2

P(g)=0.1。

The ternary character model may be following appearance:

P(e)=0.1

P(x|e)=1.0

P(p|ex)=1.0

P(l|xp)=1.0

P(a|pl)=1.0

P(i|la)=1.0

P(n|ai)=1.0

P(i|in)=1.0

P(n|ni)=1.0

P(g|in)=1.0。

● the sequence of a meta-model-under the prerequisite of not considering linguistic context in language is implemented to distribute, and each sequence is carried out inter-process as primary entity.

For instance, if the set in the training is " the dog chased the cat ", then a Dui Ying gram language model may be:

P(the)->0.4

P(dog)->0.2

P(chased)->0.2

P(cat)->0.2。

Linguistic context evidence model M _ContextComprise:

● candidate's model-under the given situation that special candidate's lexical or textual analysis arranged, implementation condition distributes on the linguistic context observed value.

● series model-under the situation of given special context, implementation condition distributes on the sequence in language or language set.In Fig. 2, series model is illustrated as language model.This language model comprises one group in a preferred embodiment corresponding to the language model of different language, for example, and LM _French, LM _German, LM _EnglishDeng.

Input evidence model M _InputComprise:

Candidate's model-under the situation of given special candidate's lexical or textual analysis, implementation condition distributes on the input observed value.

Series model-under the situation of given intended target sequence, implementation condition distributes on the candidate in language or language set.This model is illustrated as " prefix matching model " in Fig. 2.

Comprising each model of target sequence priority model R can be according to circumstances upgrade with the text of user's input.By using the dynamic language model, can calculate to a nicety the more text sequence of designated user expection of native system.

Each model has been exported one group of sequence S _R, S _Context, S _InputAnd the dependent probability that input is used as probability maker PG is estimated P _R, P _Context, P _InputProbability maker PG estimates P with the probability of above-mentioned model output _R, P _Context, P _InputCombine, generate one group of ultimate sequence prediction S _FProbable value P _F

Ultimate sequence prediction S _FCan will be shown to the user by user interface, and browse and select for the user, or can use this ultimate sequence prediction S by native system _FWith the input text of righting the wrong automatically.In case automatically or by the user selected prediction, this input text preferably is added in order to generate the linguistic context evidence of next prediction.On the contrary, the multiword symbol adds next input relevant with current entry if the user is by importing more, and then this input text preferably is added into the input evidence is distributed to prediction with modification the prior probability of working as.

The special system of introducing in detail below in the present embodiment is how to be generated by Fundamentals of Mathematics.

The instantiation expression formula (5) of bringing two evidence sources into generates:

$\frac{P\left(s\right|R)P\left(\mathrm{context}\right|s,M_{\mathrm{context}})P\left(\mathrm{input}\right|s,M_{\mathrm{input}})}{Z}---\left(12\right)$

Wherein, the Z=normaliztion constant approximates:

${\mathrm{\Σ}}_{j=1}^{\left|s\right|}P\left(s_j\right|R)P\left(\mathrm{context}\right|s_j,M_{\mathrm{context}})P\left(\mathrm{input}\right|s_j,M_{\mathrm{input}})---\left(13\right)$

As described below, this approximate value is applied to native system.The function z of an arrangement set T of let us imagination, for example:

$z\left(T\right)={\mathrm{\Σ}}_{j=1}^{\left|T\right|}P\left(s_j\right|R)P\left(\mathrm{context}\right|s_j,M_{\mathrm{context}})P\left(\mathrm{input}\right|s_j,M_{\mathrm{input}})---\left(14\right)$

It is as follows to obtain z:

Z=z(T)+z({u})*k （15）

Wherein u represents one " the unknown " sequence, and k is | the appraisal of S|-|T|, wherein | S| is the sequence quantity in the set of all possible target sequence, and | T| is the quantity of the sequence of the type estimated as " known " that at least one bottom evidence model has.Each independent evidence condition model M can return P (e|u, appraisal M), that is: under the situation of given " the unknown " sequence, the distribution on the evidence observed value.This means that substantially each evidence condition model is smoothly responsible to the distribution of himself, but must be with relevant with the proportional k of " the unknown " sequence overall estimate quantity.In fact, each model all can be understood arrangement set S ', wherein

, and P (e|s, appraisal M) will keep constant and equal P (e|u, M), although

This characteristic smoothly be that native system consider to change the employed a kind of means of level of trust in the model relevant with each evidence source.

According to expression formula (12) and (14), (s ∈ S|e M), calculates following appraisal: the preferential P of target sequence (s|R) for the conditional probability P in definite above-mentioned special system example; Linguistic context probability P (context|s, M _Context) and input probability P (input|s, M _Input).These appraisals will be discussed below and how calculate these appraisals.

Target sequence is preferential

Preferred calculating target sequence is preferentially as follows:

$P\left(s\right|R)=\left\{\begin{array}{cc}P\left(s\right|R_{\mathrm{unigram}})& \mathrm{if}(s\∈V)\\ P\left(s\right|R_{\mathrm{character}})& \mathrm{otherwise}\end{array}\right.$

Wherein V is included in R _UnigramIn arrangement set, the enforcement of model then realizes according to the known technology that makes up a level and smooth gram language model and level and smooth Markov chain character model.Some application technologies that are used for these models of enforcement are listed hereinafter.But the technology that other are fit to is also unlisted.

● level and smooth n unit's entry or character model (known in this field);

● as the patent documentation of＜GB Patent Application No. 0917753.6〉in the multilingual model of self-adaptation put down in writing;

● as＜document: Scheffler2008〉the middle PPM(prediction by partial matching that puts down in writing, partial match estimation) language model;

● under certain probability by the morphological analysis engine that constitutes the morphological component formation sequence.

By comprising target sequence priority model R, native system has improved the expected sequence accuracy of predicting.In addition, target sequence priority model R can realize the reasoning based on character of invisible (unseen) target sequence, that is to say, native system can infer better that the unknown object sequence is with all possible target sequence of general estimation.

The linguistic context probability

Preferably by second candidate's model evaluation linguistic context probability P (context|s, the M _Context) to propose following expression formula (16).Although this mode is the preferred means of assessment probability, the present invention is not limited to make the probability of assessing in this way.

$P\left(\mathrm{context}\right|s,M_{\mathrm{context}})=\frac{{\mathrm{\Σ}}_{j=1}^{K}P\left(s\right|c_j,M_{\mathrm{context}-\mathrm{sequence}})P\left(c_j\right|\mathrm{context},M_{\mathrm{context}-\mathrm{candidate}})P\left(\mathrm{context}\right|M_{\mathrm{context}})}{P\left(s\right|M_{\mathrm{context}})}---\left(16\right)$

Therefore, in order to determine the linguistic context probability, be calculated as follows every: the linguistic context sequence is estimated P (s|c _j, M _{Context-sequence}); Linguistic context candidate estimates P (c _j| context, M _{Context-candidate}); Linguistic context is preferentially estimated P (context|M _Context); And target sequence is preferentially estimated P (s|M _Context).These appraisals will be discussed below and how calculate these appraisals.

The linguistic context sequence is estimated

At given particular candidate sequence c _jSituation under, the linguistic context sequence is estimated P (s|c _j, M _{Context-sequence}) be the probability of the target sequence s under the linguistic context series model.The linguistic context series model is a kind of function that returns the probability of target sequence under the situation of given linguistic context sequence, i.e. f in form _s(t _Target, t _Context)=P (t _Target| t _Context, θ _s), wherein, θ _sIt is the parameter of model.Therefore, being calculated as of linguistic context sequence probability: P (s|c _i, S)=f _s(s, c _i).Can use multiple different technologies to calculate this appraisal.For example, the level and smooth frequency analysis on the linguistic context training data, similar with equation (21) and to estimate relevant mode and explain with target sequence is preferential.Alternatively, can separately or be used in combination following any:

● n gram language model (known in this field);

● as the patent documentation of＜GB Patent Application No. 0917753.6〉in the multilingual model of self-adaptation put down in writing;

● as＜document: Scheffler2008〉the middle PPM(prediction by partial matching that puts down in writing, partial match estimation) language model;

● the HMM(Hidden Markov Model of generation, hidden Markov model) probability part-of-speech tagging device＜reference: 008.LingPipe4.1.0.http: //alias-i.com/lingpipe (accessed September26,2011) or Thede, S.M., Harper, M.P., 1999 〉;

● be used for the natural language resolver of the probability of returning part sentence, as RASP＜reference: Briscoe, E., J.Carroll and R.Watson2006 〉;

● admit to represent the input characteristics of linguistic context sequence and target sequence and the neural network of output probability (techniques well known).

Native system is not limited to above-mentioned technology, and any technology that can be used for calculating the linguistic context sequence probability all is applicable to native system.

As indicated above, M _{Context-sequence}Can comprise a plurality of language models corresponding to multiple different language.In order to determine the conditional probability of expression formula (16), can use the language model relevant with entry to determine this conditional probability.As a kind of explanation, mentioned in the above-mentioned example from English model (LM _English) and French model (LM _French) the middle prediction entry of extracting out " pain ".In this case, expression formula (16) is by P (context|pain, LM _English) andP (context|pain, LM _French) determine, wherein from French model (LM _French) in " Pain " that extract out be different from from English model (LM _English) middle " Pain " that extracts out, although should predict identical lexically.By entry is associated with its source model, native system has been simplified the processing mode of the identical entry of morphology, because native system has only kept the most probable entry in the identical entry of two or more morphology.In addition, native system has been simplified the conditional probability calculating of expression formula (16).This simplification is feasible, although because the morphology of entry is identical, entry has the different meaning of a word in different language, can treat this entry with a certain discrimination thus.

Transfer back to Fig. 2 like this, by model M _ContextThe entry S set that generates _ContextCan comprise M _ContextIn the entry of any language model (or candidate's model).

Linguistic context candidate estimates

Linguistic context candidate estimates P (c _j| context, M _{Context-candidate}) be a kind of function, its form is f _{Context-candidate}(t)=P (t| θ _{Context-candidate}), wherein t is arbitrary sequence, θ _{Context-candidate}Be model parameter.What like this, linguistic context candidate condition was estimated is calculated as follows: P (c _j| context, M _{Context-candidate})=f _{Context-candidate}(c _j).

In an optimum decision system, linguistic context candidate is sequence, and the linguistic context candidate collection is expressed as directed acyclic graph (DAG, directed acyclic graph), and each node in the directed acyclic graph comprises the subsequence that contains one or more characters.Each border is assigned probability, and in a preferred embodiment, directed acyclic graph preferably also has the specific properties that each path is restricted to same length.In this article, the variant of directed acyclic graph is called as probabilistic limited sequence chart (PCSG, probabilistic constrained sequence graph).Each independent candidate sequence then is expressed as unique path by PCSG, and linguistic context candidate pattern function is calculated as the probability of the delegated path of linguistic context candidate model for given candidate's rreturn value.

In form, PCSG comprises the four-tuple that contains a group node N, root node r, one group of directed edge E and one group of parameter (probability) θ:

G=(N,r,E,θ) (17)

Boundary representation between two node n, the n ' is that (n → n'), the probability tables of shifting to n ' from n along the border is shown P (n'|n).Path through G originates in node r, and follows one from the outwardly directed border extension of each node of visiting, till arriving at the node that does not contain the border of going out (outgoing edge).The attribute of G is as follows:

1) G is directed acyclic graph (DAG);

2)

That is: the node of all except root node must have at least one and enters border (incoming edge);

3） $\∃m,k\∈N.\∀n\∈N.(m\→n)\∈E\⇒(n\→k)\∈E,$ That is: add follow-up common path immediately again from the path that given node is told.This attribute has strictly limited the structure of this figure, and has hinted that all paths have identical length, has reduced the normalization needs that path probability is calculated.

Linguistic context candidate's pattern function calculates the probability following (being equal to linguistic context candidate estimation) of given path:

P(c _j|context,M _{context-candidate})=f _{context-candidate}(c _j)=P(p _j|G) （18）

Wherein, P (p _j| G) be path probability, be calculated as the product on each border in the path:

$P\left(p_{j}G\right)=P\left(n_1\right|r){\mathrm{\Π}}_{k=2}^{K}P\left(n_k\right|n_{k=1})---\left(19\right)$

Wherein, K is the boundary number in the path.It should be noted that this preferred formula is equivalent to internodal implicit independence assumption.This is that the sequence probability of candidate sequence is not modeled because in this case, and the variation probability among the candidate is modeled.Therefore, following Column Properties has kept borderline probability:

$\∀n\∈N.{\mathrm{\Σ}}_{(n\→m)\∈E}P\left(m\right|n)=1---\left(20\right)$

That is to say that all borderline probability sums one of going out that come from given node n are decided to be 1.This means that also following expression formula is effective: That is: the probability sum in all paths equals 1 among the PCSG.

A certain example is illustrated these concepts with help.Consider following 12 linguistic context candidate sequences:

·“Sunday at3pm” ·“sunday at3pm” ·“Sun at3pm”

·“Sunday at3pm” ·“sunday at3pm” ·“Sun at3pm”

·“Sunday at3p.m.” ·“sunday at3p.m.” ·“Sun at3p.m.”

·“Sunday at3p.m.” ·“sunday at3p.m.” ·“Sun at3p.m.”

These linguistic context candidate sequences can be represented by ' | ' by clear and definite word circle of following PCSG(, and empty sequence is by " φ " expression) be expressed as:

According to linguistic context candidate model and adopt expression formula (19), give the border with probability assignments, for example:

Then, the candidate's probability that from PCSG, generates above-mentioned 12 sequences following (to be concise in expression clearly in order making, only to have listed three examples):

P("sunday at3pm"|"sunday at3pm",C)=0.6*1.0*1.0*0.6*1.0*0.7=0.252

P("Sunday at3pm"|"sunday at3pm",C)=0.3*1.0*1.0*0.4*1.0*0.7=0.084

P("sun at3p.m."|"sunday at3pm",C)=0.1*1.0*1.0*0.4*1.0*0.3=0.012

Be used for to make up DAG and change to the particular example that the model detail of node allocation probability will depend on native system.Above-mentioned graphic examples to three kinds of general variations are encoded:

● the borderline branch of word (the single meaning of possibility);

● the branch in capital and small letter (case) variation;

● the branch on morphology changes.

Be understood that the variation of any kind all can be encoded in this framework.Scheme before another example will turn to, for example, if native system dopes " on " and " in ", and the user has selected " in ", then it can be encoded into the branch that also has the small probability of distributing to " on " the probability right of distributing to " in " except having, with the possibility of the unexpected acceptance error suggestion of expression user.In these cases, enroll following principle:

● be lower than " Sun " of abbreviated form with the possibility of ' sunday ' of lowercase ' s ' beginning, and the possibility of " Sun " of abbreviated form is lower than full spelling distortion ' Sunday ';

● the possibility of the participle situation (tokenization case) that " pm " and numeral " 3 " are split slightly is lower than the not situation of participle;

● the possibility of fullstop distortion " p.m. " is a shade below no fullstop form " pm ".

In the following manner, preferably made up the particular example of linguistic context candidate PCSG at algorithm by homing sequence s:

1) by s being encapsulated among the node ns that is connected in root node, converts s to PCSG;

2) by introducing the branch node on the change point, iteration is torn open and is analysed ns.

For instance, consider PCSG construction algorithm that original series " sunday at3pm " is worked.At first, step 1:

Native system is disposed probability participle device, and the result is as follows:

It should be noted that, because the attribute 3 of above-mentioned PCSG, modification will always show as the structure infix form of branch-also-reclosing (branch-and-rejoin), for special circumstances, modification has a node branch, this is convenient to subsequent treatment very much, because it can not have influence on the probability in whole paths.To introduce in further detail hereinafter according to model and add boarder probability.Continue to introduce this algorithm, capital and small letter (case) deformation analysis device is disposed as follows:

At last, morphology deformation analysis device is disposed as follows:

It should be noted that the attribute 3 because of PCSG, each branch must converge before the fork again.This means, in some cases, if occur two take-off points continuously, then must insert empty node.

Boarder probability preferably is assigned to PCSG.About the parameter of linguistic context candidate model, can preferably implement boarder probability and distribute.Visual interpretation for these probability has 2 points:

1) the probability appraisal of user's expected sequence of specific branch is distributed in their expressions.For example, if the user imports " Dear ben ", we perhaps can think and may they in fact want input " Dear Ben ";

2) their expressions are the compensation probability of effective spelling distortion of the sequence that observes for specific branch.For example, if user's input " See you on Thur ", then the selectable correct orthographic form of " Thur " can be " Thurs ".

Under the situation of given a certain background model information, the probability of distributing to specific border also can be subjected to correctly spelling the influence of the estimated probability of variant.For example, in fact can reuse the probability appraisal that linguistic context series model s tries to achieve different correct spelling variants.This probability is estimated and can be used in combination with other probabilistic quantities, generates branch's probability.The context of use series model means the actual linguistic context series model S that comprises of linguistic context candidate MODEL C by this way, obviously runs counter to the independence assumption (attribute 7 above) between candidate's model and the series model thus.Yet this hypothesis will not appear under the linguistic context situation, and is therefore comparatively safe.

Following example will help explanation.In a preferred embodiment, suppose that linguistic context candidate model uses following algorithm assigns probability:

1) sequence that observes obtains probability 0.8, and other sequences are evenly divided remainder equally;

2) estimate measurement numerical value by the linguistic context series model;

3) normalization numerical value makes it satisfy above-mentioned PCSG attribute 19.

According to above-mentioned PCSG example, following branching into:

Because " sunday " is raw observation, at first the step 1 by above-mentioned algorithm is its allocation probability value 0.8, and other borders then respectively are assigned probable value 0.1.The appraisal of being returned by the linguistic context series model in this example is as follows:

P("sunday"|C ^S)=0.01

P("Sunday"|C ^S)=0.04

P("Sun"|C ^S)=0.02

Wherein, C ^SBe illustrated in linguistic context candidate model context of use series model in this case.Therefore, in this example, the probability (difference) of distributing to the not normalization on each border and normalization (through rounding up) is as follows:

Linguistic context is preferentially estimated

By the frequency of the normalization initiation sequence t relevant with linguistic context, can roughly estimate linguistic context and preferentially estimate P (context|M _Context)

$P\left(\mathrm{context}\right|M_{\mathrm{context}})\≅\frac{\mathrm{freq}\left(t\right)}{E_{t^{\′}}\mathrm{freq}\left(t^{\′}\right)}---\left(21\right)$

Wherein freq (t) is the frequency of sequence t in the training data, and denominator is the frequency sum of all sequences in the training data.Sequence " t " in the expression formula (21) is the current linguistic context as the native system input.Preferentially according to the probable value of probability weight prediction, the corresponding model of wherein extracting this prediction out comprises the corresponding probability of given linguistic context sequence to linguistic context.In order to realize this process, linguistic context is preferentially according to the appraisal weight estimation value of expression formula (21).

In actual applications, for example can by settle in invisible sequence hypothesis (occurrence assumption) appears or be retracted into all sequences all limited (low level) in the sightless example estimate smoothly this appraisal.For instance, if linguistic context is ternary model, then prediction engine can be return and be constituted binary or monobasic appraisal.

Linguistic context preferentially provides dual function: help the normalization probability to estimate; And when can't providing useful information, context model provides simple " model detection ".Can't provide information (for example, when last entry is unknown) if the linguistic context sequence is estimated for the N meta-model, linguistic context is preferentially estimated can use most probable linguistic context this model of weighting in large quantities, and the prediction of this model is promoted on the prediction of other models." most probable linguistic context " is to estimate (21) maximal value on a plurality of model sets, for example language model set LM _English, LM _French, LM _GermanFor instance, if linguistic context is " The dog chased ", then can expect that with appearing to compare in the French this linguistic context more likely appears in the English.Therefore, the conditional probability of expression formula (21) is for LM _EnglishTo be maximum, and the probability maker thus in a large number weighting from LM _EnglishPrediction probable value but not from LM _FrenchThe probable value of prediction, therefore, LM _EnglishMore be subjected to preferential " preference " estimated of linguistic context.

Therefore, consider linguistic context, linguistic context is preferentially estimated a large amount of weightings from the only language model in a plurality of language models relevant with multilingual.Like this, the preferential appraisal of linguistic context can detect the affiliated language of text that someone imports.

Target sequence is preferentially estimated

Can be according to preferentially estimating with linguistic context, the same mode of expression formula (21) is used through level and smooth training data frequency analysis estimation target sequence and is preferentially estimated P (s|M _Context), that is: can be preferential by the target sequence frequence estimation target sequence on all sequences in the normalization linguistic context training data:

$P\left(s\right|M_{\mathrm{context}})\≅\frac{\mathrm{freq}\left(s\right)}{{\mathrm{\Σ}}_{s^{\′}}\mathrm{freq}\left(s^{\′}\right)}$

Wherein freq (s) is the frequency of target sequence in the training data, and denominator is whole frequency sums of all target sequences in the training data.Denominator roughly is equivalent to entry (the comprising the dittograph bar) sum in the training data.

Input probability

Utilize first candidate's model assessment input probability P (input|s, the M _Input):

$P\left(\mathrm{input}\right|s,M_{\mathrm{input}})={\mathrm{\Σ}}_{j=1}^{K}P\left(\mathrm{input}\right|c_j,M_{\mathrm{input}-\mathrm{candidate}})P\left(c_j\right|s,M_{\mathrm{input}-\mathrm{sequence}})---\left(22\right)$

Therefore, in order to determine input probability, need to calculate following appraisal: the input candidate estimates P (input|c _j, M _{Input-candidate}) and list entries appraisal P (c _j| s, M _{Input-sequence}).Introduce this two kinds of appraisals below.

Input the candidate estimate

The input candidate estimates P (input|c _j, M _{Input-candidate}) incoming event and the function on the sequence: the f that are defined as observing _{Input-candidate}(i, t)=P (i|t, θ _{Input-candidate}), θ wherein _{Input-candidate}It is the parameter of model.Input observed reading i encodes in list entries expected structure (ISIS, input sequence intention structure) arbitrarily.This list entries expected structure is the ordered list with the arrangement set of probability mapping:

{(t ₁₁→P(i ₁|t ₁₁),(t ₁₂→P(i ₁|t ₁₂),...},{(t ₂₁→P(i ₂|t ₂₁),(t ₂₂→P(i ₂|t ₂₂),...},...

It should be noted that each appraisal has form P (i _j| t _Jk), if that is: the user has planned list entries t _Jk, we should observe incoming event i _jProbability what are.Consider following ISIS example:

$\left[\begin{array}{c}\{(H\→0.5),(h\→0.3),(g\→0.1),(j\→0.1)\}\\ \{(e\→0.8),(w\→0.1),(r\→0.1)\}\end{array}\right]$

This ISIS example is encoded for a kind of like this scheme, and whether native system estimation user plans to import the character ' H ' of for example following character ' e ' thereafter in this scheme, estimates that like this incoming event that observes has probability 0.5 and 0.8 respectively.

The method that generates these probability distribution is not theme of the present invention.Or rather, given prominence to a series of suitable technology, for instance:

-can distribute according to the character generating probability of given target keywords on the keyboard layout, for example qwerty keyboard if the user knocks the zone corresponding with " H " key, then may comprise the character " G " and " J " that have certain probability in ISIS.

-can distribute according to the distance between touch coordinate (touch-screen dummy keyboard) and the specified button coordinate (or certain distance function, for example square etc.) generating probability.

In a preferred systems, the input candidate is sequence, and the input candidate collection is expressed as the PCSG(EPCSG of expansion).EPCSG is a kind of PCSG, but has the additional structure of running counter to Standard PC SG attribute (definition hereinafter).As the linguistic context situation, represent each candidate sequence by unique path of passing EPCSG, and given candidate's input candidate pattern function rreturn value is calculated as the normalization probability of its delegated path.

Input candidate EPCSG generative process begins with the ordered list of native system by the sequence probability pair set that generates with user interactions, wherein the probability distribution in each subclass representative of consumer list entries expection.

The algorithm that is generated input candidate EPCSG by input ISIS comprised for two steps:

1) converts ISIS to PCSG;

2) insert additional generalized structure (generalizing structure), thereby generate EPCSG.

Step 1 is flat-footed.Root node with new PCSG begins, and this algorithm is each the structure branch that distributes among the ISIS.Above-mentioned ISIS result in the step 1 is as follows:

Step 2 is used two existing PCSG of additional structure finishing.First structure is empty node subpath (its suitable PCSG framework), and second structure is ' asterisk wildcard ' structure (converting PCSG to EPCSG).The application example of step 2 is as follows:

The symbol of asterisk wildcard (being expressed as ' * ') is actually the simple expression way of the branch that comprises/generate each symbol in the character set.The asterisk wildcard structure is a limited circulation, has therefore run counter to the acyclic attribute of Standard PC SG.The EPCSG expansion allows only to use the asterisk wildcard circulation at convergence point, and numerical value e, w are the probability constants of predesignating.It should be noted that in this case each take-off point has sky node additional (being two in this case), and each convergence point has asterisk wildcard additional (being in this case).These generalized structures can be omitted the one or more characters in the target sequence (having asterisk wildcard probability w) or insert one or more error characters (having hole node probability e) with respect to the user.Be understood that how structure that these are the extra details that adds PCSG can change according to the different instances along with native system such as computational resource, series model length.

Empty node subpath makes native system can abandon the character of user error input, and not so this error character can cause incorrect chain through PCGS.

Rely on these additional generalized structures (especially asterisk wildcard branch), the quantity through the path of PCSG is increased sharply.For example, suppose that size is 50 character set, 1020 different paths can be arranged through the PCSG of above-mentioned simplification.For the ISIS of reality, there are tens of very hundreds and thousands of different paths.This optimum decision system preferably uses this combination of following technical finesse to increase sharply in mode alone or in combination.

● use trie (word lookup tree, techniques well known) to neglect and predict that those are not the paths of sequence prefix in the vocabulary;

● the probability of use threshold values is deleted those unlikely relatively paths.Threshold values is set to the ratio of the differential of the lower sequence of current most probable sequence and possibility.Given threshold values t, the path L of current investigation.If the maintenance following relationship is then deleted path n ₁N _L:

$\frac{P\left(n_1\right|r){\mathrm{\&Pi;}}_{j=2}^{L}P\left(n_j\right|n_{j-1})}{\arg {\max }_m\&lsqb;P\left(m_1\right|r){\mathrm{\&Pi;}}_{j=2}^{L}P\left(m_j\right|m_{j-1})\&rsqb;}<t---\left(23\right)$

● list entries model T is used to the probability threshold values equally.Given difference or limited threshold values t, and be L:{c by length ₁..., c _KThe arrangement set that forms of all paths.If maintenance following relationship, then deleted representation particular sequence c _PGiven path p:

$\frac{P\left(c_p\right|T)}{\arg {\max }_j\&lsqb;P\left(c_j\right|T)\&rsqb;}<t---\left(24\right)$

Can also dispose separately or with above-mentioned wherein one or all technical combinations dispose the technology that other are applicable to that treatment combination increases sharply.

List entries is estimated

Under the situation of given target sequence, list entries is estimated P (c _j| s, M _{Input-sequence}) be the distribution on the candidate sequence, and can be estimated as normalized target function:

$P\left(c_j\right|s,M_{\mathrm{input}-\mathrm{sequence}})=\frac{\mathrm{\&delta;}(s,c_j)}{z}---\left(25\right)$

Wherein, if t ' is the prefix of t, then δ (t, t')=1, otherwise δ (t, t')=0, and the z=∑ _kδ (s, c _k), i.e. all candidate's sums.

It should be noted that if present candidate's uniqueness, and allow candidate collection to comprise all possible sequence, then can recomputate normalization factor: Z=length (s).For instance, given target sequence " the " always just in time has three matching candidates: " t ", " the " and " the ".

Therefore, the invention provides a kind of text prediction engine and system of routine, and the particular example of text prediction engine or system, text prediction engine or system can generate one group and have dependent probability value P respectively _FSequence prediction S _F

The present invention also provides a kind of correlation method for the treatment of the user version input.Be back to Fig. 1 and said system, this method comprises that receiving the text that inputs to such as user interfaces such as electronic equipments imports; Use the prediction of text prediction engine formation sequence SF and relevant probable value PF; And this sequence prediction provided to described user interface.

As the introduction for native system, conventional method comprises: by comprising the prediction of text prediction engine formation sequence and the dependent probability value of one or more models.In a preferred embodiment, this method comprises: use at least one evidence source e by target priority model R and at least one ₁, e ₂... the model M of generation forecast ₁, M ₂... the formation sequence prediction.As mentioned above, about this system, and especially expression formula (12) is to (15), and this method comprises: the probable value normalization factor generation normalization probable value of being obtained by the constant sum of the possible sequence prediction of the probable value of n most probable sequence prediction and expression residue by estimation.

With reference to Fig. 2, in above preferred embodiment, final prediction sets S _FAnd relevant probable value P _FBy probability maker PG according to respectively from target priority model R, context model M _ContextWith input model M _InputThe prediction sets S that extracts out _R, S _Context, S _InputGenerate.In this embodiment, the linguistic context of user input sequence is used as from context model M _ContextIn extract the evidence of prediction out, and the relevant user input sequence of the current word of attempting importing with the user is used as from input model M _InputThe middle evidence of extracting prediction out.

Other aspects and the said system of this method are similar, for example, and in a certain embodiment of this method, if the corresponding probable value of sequence prediction then only offers user interface with these sequence predictions respectively more than or equal to first threshold values.

As mentioned above, related to one and implement the system of generalized structure to determine that the linguistic context candidate estimates in PCSG, in a preferred embodiment of this method, at least one group of sequence prediction is equivalent to be inputed to by the user adjustment or the invulnerable release of the text input of user interface.

Description according to said system is analogized, and can determine other aspects of the inventive method at an easy rate.

Following is claims of the non-exhaustive nothing left of above-described embodiment of stating in this application:

1. system comprises:

Text prediction engine, it comprises:

At least one is used for having from one group of evidence source generation the model of the sequence of dependent probability appraisal;

Be used for generating one group of model that has the sequence of dependent probability appraisal; And

The probability maker, it is used for receiving the sequence of respectively organizing that has the dependent probability appraisal and also generates the sequence prediction that has the dependent probability value.

2. the system of embodiment 1 comprises a plurality of for generated a plurality of models that have the arrangement set of dependent probability appraisal by some evidence sources.

3. in the system of embodiment 1 or 2, the probability maker is used for generating one group of sequence prediction according to an any amount corroboration source.

4. in the system of embodiment 3, one of them evidence source comprises user input text.Modes such as this user input text can be selected by the user, character input, speech recognition are imported.

5. system comprises:

User interface is used for receiving user input text;

According to the text prediction engine of first aspect or other suitable text prediction engine arbitrarily, be used for receiving the text input that comes from user interface and generate the sequence prediction that has the dependent probability value.

6. in the system of embodiment 5, text prediction engine comprises:

Context model is used for receiving the text of being imported by the user and generates one group of sequence and relevant probability appraisal;

Input model is used for receiving the text of being imported by the user and generates one group of sequence and relevant probability appraisal;

One model is used for generating one group of sequence and estimates with relevant probability; And

The probability maker is used for receiving from above-mentioned a plurality of models and respectively organizes sequence and relevant probability appraisal and generate one group of sequence prediction and relevant probable value.

7. in the system of embodiment 6, described user interface is used for the above-mentioned sequence prediction of demonstration and selects for the user.

8. in the system of embodiment 7, this system is described sequence prediction ordering according to described probable value and described sequence prediction is shown as ordered set.

9. in embodiment 6 to 8 in the system of any one embodiment, wherein, this system preferably uses the automatic correction of described sequence prediction to input to the wrong input text of user interface.

10. in embodiment 6 to 8 in the system of any one embodiment, described text prediction engine is used for generating based on the sequence prediction in a corroboration source arbitrarily.

11. the method for a process user text input comprises:

Reception inputs to the text of user interface;

Utilize the prediction of prediction engine formation sequence and relevant probable value;

This sequence prediction is offered described user interface.

12. in the method for embodiment 11, comprising: described sequence prediction is presented at described user interface selects for the user.

13. in the method for embodiment 12, comprising: according to the dependent probability value of described sequence prediction, be described sequence prediction ordering.

14. in the method for embodiment 13, comprising: show that sorted sequence prediction selects for the user.

15. in embodiment 11 to 14, in the method for any one embodiment, comprise and utilize the automatic correction of described sequence prediction to input to the step of the wrong input text of user interface.

16. in the text prediction engine of in aforementioned any one embodiment, using, comprising: context model, input model, target priority model and probability maker.

17. in the text prediction engine of embodiment 16, described context model receives the text of being imported by the user with described input model user and generates one group of sequence and relevant probability appraisal.

18. in the text prediction engine of embodiment 16 or 17, described target priority model is used for generating one group of sequence to be estimated with relevant probability.

19. in the text prediction engine of an embodiment in embodiment 16 to 18, described probability maker is used for respectively organizing sequence and relevant probability appraisal from described model reception, and generates each to one group of sequence prediction of probable value should be arranged.

The above only is preferred embodiment of the present invention, and is within the spirit and principles in the present invention all, any modification of doing, is equal to replacement etc., all should be included within protection scope of the present invention.

Claims (36)

1. text prediction engine comprises:

At least one model, it is used for having first group of sequence that dependent probability is estimated from the generation of evidence source;

The probability maker, it is used for receiving the described first group of sequence that has the dependent probability appraisal and generates one group of sequence prediction that has the dependent probability value;

Wherein, received under the situation of all possible sequence the described probable value of normalization on all possible sequence prediction that is generated by described probability maker by described probability maker given.

2. text prediction engine according to claim 1 is characterized in that, also comprises priority model, and it is used for generating the second group of sequence that has the dependent probability appraisal.

3. text prediction engine according to claim 1 and 2 is characterized in that, described model generates first group of sequence according to the uncertainty in described evidence source and the described evidence source.

4. according to claim 2 or 3 described text prediction engine, it is characterized in that described probability maker receives and has described first, second group sequence that dependent probability is estimated.

5. according to any described text prediction engine in the claim 1 to 4, it is characterized in that, described probability maker is estimated the normalization factor of described probable value by n most probable sequence prediction is represented the constant addition with remaining probable value that may sequence prediction.

6. text prediction engine according to claim 5 is characterized in that, described constant is represented the probable value by the remaining possibility sequence prediction of described model and the generation of described priority model.

7. according to any described text prediction engine in the claim 1 to 6, it is characterized in that described model comprises a plurality of models that have a plurality of first group of sequence of dependent probability appraisal for generation.

8. text prediction engine according to claim 7 is characterized in that, described a plurality of models generate a plurality of first group of sequence according to a plurality of evidences source.

9. text prediction engine according to claim 8 is characterized in that, one in described a plurality of evidences source comprises user input text.

10. text prediction engine according to claim 9 is characterized in that, described text prediction engine is the part of a certain system, and described user input text by one or more users select, character input or speech recognition be input in this system.

11., it is characterized in that described text prediction engine comprises given linguistic context sequence according to corresponding model probability is weighted the probable value of described sequence prediction according to claim 7 or 8 described text prediction engine.

12. text prediction engine according to claim 11 is characterized in that, described a plurality of models comprise a plurality of language models corresponding with multiple different language; And the probable value of the described text prediction engine pair sequence prediction corresponding with the language model of the most probable language that relates to user input text is carried out five-star weighting.

13., it is characterized in that each evidence source is moulded by the corresponding model that is used for generating the sequence that has the dependent probability appraisal according to any described text prediction engine in the claim 10 to 12.

14. text prediction engine according to claim 13 is characterized in that, under the situation of given described sequence prediction, described probability maker with each evidence source as other on evidence the separate component of having ready conditions in source handle.

15. according to aforementioned any described text prediction engine of claim, it is characterized in that, described model comprises context model and input model, and described context model is used for receiving the text of user's input and generates one group of sequence with described input model to be estimated with relevant probability; And

Described priority model comprises for generating the target priority model that one group of sequence and dependent probability are estimated.

16. text prediction engine according to claim 15 is characterized in that, described input model comprises candidate's model and language model.

17., it is characterized in that described context model comprises candidate's model and prefix matching model according to claim 15 or 16 described text prediction engine.

18. according to any described text prediction engine in the claim 15 to 17, it is characterized in that described target priority model comprises character model and a meta-model.

19. the description of aforementioned texts prediction engine is with reference to accompanying drawing and as shown in drawings.

20. a system comprises:

User interface, it is used for receiving the text by user's input;

Text prediction engine, it be used for to receive from the described text of described user interface input and generates one group of sequence prediction that has the dependent probability value, wherein, all possible sequence prediction on the described probable value of normalization;

Wherein, described text prediction engine also is used for providing described sequence prediction to described user interface.

21. system according to claim 20 is characterized in that, described text prediction engine is as any described prediction engine in the claim 1 to 19.

22., it is characterized in that only having one during more than or equal to the probable value of the first probability threshold values when described sequence prediction according to claim 20 or 21 described systems, described text prediction engine provides sequence prediction to described user interface.

23., it is characterized in that if described text prediction engine has the corresponding probable value that is equal to or greater than the second probability threshold values, then described text prediction engine is from the list entries prediction of the described system of trend according to any described system in the claim 20 to 22.

24. according to any described system in the claim 20 to 23, it is characterized in that described system is presented at described sequence prediction on the described user interface and selects for the user.

25. system according to claim 24 is characterized in that, described probability maker sorts to described sequence prediction according to the dependent probability value of described sequence prediction, and described user interface shows described sequence prediction as ordered set.

26. the description of aforementioned system is with reference to accompanying drawing and as shown in drawings.

27. the method for a process user input text comprises:

Reception inputs to the text of user interface;

Use text prediction engine to generate one group of sequence prediction and relevant probable value, wherein, the described probable value of normalization on all possible sequence prediction;

Described sequence prediction is offered described user interface.

28. method according to claim 27, it is characterized in that, the step that generates the normalization probable value comprises: by the probable value of n most probable sequence prediction is represented the constant addition with remaining probable value that may sequence prediction, estimate the normalization factor of described probable value.

29. according to claim 27 or 28 described methods, it is characterized in that, also comprise: described sequence prediction is presented on the described user interface selects for the user.

30. method according to claim 29 is characterized in that, by described text prediction engine described sequence prediction is sorted, and shows in order for described user interface.

31. according to any described method in the claim 29 to 30, it is characterized in that, only when the corresponding probable value of described sequence prediction during more than or equal to first threshold values, described sequence prediction offered described user interface.

32. according to any described method in the claim 29 to 31, it is characterized in that at least one in the described sequence prediction is equivalent to be inputed to by the user adjustment or the invulnerable release of the text of described user interface.

33. according to any described method in the claim 29,30 and 32, it is characterized in that, also comprise: input has the sequence prediction greater than second threshold values or the probable value on second threshold values automatically.

34., it is characterized in that described probability maker comprises for a plurality of models that generate one group of sequence prediction and dependent probability value according to any described method in the claim 27 to 33; Described probable value comprises that according to corresponding model the probability of given linguistic context sequence is weighted.

35. the description of preceding method is with reference to accompanying drawing and as shown in drawings.

36. a computer program comprises: computer-readable medium wherein stores for making the processor enforcement of rights require the computer program of 27 to 35 any described methods.

Images